GPS TEC calibration: details and practical...

GPS TEC calibration: details and practical aspectsL. Ciraolo

IFAC-CNR, Firenze / ICTP, Trieste

E-mail: l.ciraolo@ifac.cnr.it, lciraolo@ictp.it

Second Workshop on Satellite Navigation Science and Technology for Africa

6-24 April 2010

the Abdus Salam ICTP, Trieste, Italy

Geometrical Optics, Rays, Propagation DelaysPhase delay L, Optical path Λ = L · λGroup or Code Delay P = c · G, G = dL / df

Two carriersf1 (1575.42 MHz), f2 (1227.6 MHz)

Modulated by codes P and C/A

Arc, set of continuous observations

GPS observables L1, L2, P1, P2, C1

GPS scenario

GPS

Ground receiver

.

InterferenceRay

Multi-path

Thermal noise Hardware Delays (HD)

Output files

Propagation delays, Disturbances, Hardware Delays, Multi-Path

Propagation dela

y

Propagation Delays

Propagation and Atmospheric contributions to optical path Λ:

Geometric (Distance), Τropospheric, Ionospheric

Λ = D + T + I

Equivalent Group Path P = Group delay G × speed of light

P = G · c = D + T - I

Refractivity R = n -1, n Index of Refraction

2

2e

2IonoIonoatm

cf40.3TEC

cTD

dfdLG

cf40.3TECT)(D

cf

λITDL

fTEC40.3I (s)ds,NTEC

,f

N40.3R(s)dsRI(s)dsRT

++

==

−+=++

=

⋅−==

⋅−===

∫

∫∫ e

Measurements introduce additional "delays"

Hardware electronic delays originating

in satellite and receiver, β, γ

Offset (delay, ambiguity) for phase Ω

Noise n

Multipath m

User clock offset τ

Code delay affected by user clock offset is pseudorange

P = D + T - I + β + γ + n + m + τ

For following discussion, noise and multipath can be neglected for phase delays. Hardware delays for phase are included in Ω

Λ = D + T + I + Ω

τ1 = ( D+T+I1) / c

Propagation delays

Space

Modulator

δ tT1

δ tT2

L1

L2

TX Transmitter, satellite

Code Generator

Modulator

Correlator

δ τR1

δ τR2

RX Receiver

Code Generator

Correlatorτ2 = ( D+T+I2) / c

τ2 + δ tT2 + δ τR2

τ1 + δ tT1 + δ τR1

Code hardware delays

Osc 10.23 MHz

× 154

× 120

L1 = ( D+T-I1) / λ1

L2 = ( D+T-I2) / λ2

Propagation delays

δφ T1

δφ T2

÷ 154H

δφ R1δ φR2

L1

L2

TX

RXSpaceHardware delays

Hardware delays

L2 + δφ T2 + δ φR2

L1 + δ φT1 + δ φR1

Phase Hardware Delays

Availing GPS delays P1, P2, L1, L2, C1

Users aiming to determine their position, will get rid of ionospheric contribution taking proper combinations of them.

Users aiming to investigate ionosphere, will simply compute differential delays

Differential pseudorange

P2 – P1

Differential phase path

Λ1 – Λ2 = L1 ⋅ λ1 - L2 ⋅ λ2

Both differential delays are in meters.

Following steps:

Show dependence on TEC

Transform to TEC units (1016 electrons/m2 ), TECu

The differential Delays

For the carrier i (i = 1,2), contributions with no index do not depend on frequency and cancel out forming differential delays

Pi = Gi · c = D + T - Ii + βi + γi + ni + mi + τ,

ΔP = P2 – P1 = I1 – I2 + Δβ + Δγ + Δn + Δm

Λi = D + T + Ii + Ωi

ΔΛ = Λ1 – Λ2 = I1 – I2 + ΔΩ

Divide by k·10-16, drop out the Δ symbol to obtain the phase slants SP and group or code slants SC in TECu, 1 TECu = 1016 electrons/m2, disregard radio noise n

⎟⎟⎠

⎞⎜⎜⎝

⎛−== 2

12

2 f1

f1TEC40.3kTEC · k I1 – I2

β m TEC P1) -(P2 · k1 S

Ω TEC 2)Λ -1Λ·( k1 S

C

P

γ+++==

+==

The classical interpretation of TEC as the numbers of electronscontained in a column of unitary base along the ray

Rx,User

Never forget: TEC > 0

∫=Tx

Rx edsNTECTx,GPS

Note for the following: expressions for observations like

S = TEC + b

denote the set of all available observations used for performing some specific task.

Actually observations should be indexed as Sijt meaning that the individual observed quantity , the “slant ”, refers to ith satellite, jth station, tth time.

Biasing terms can still be indexed according to satellite and station (not time as assumed to be constant), but also according to the specific observed arc.

When needed for clarity, indexing will be explicitly adopted.

Plot of SC arcs for one day

* Evidence that calibration is needed: TEC is a positive quantity

Sample SC , one arc: the common situation

Sample SP , one arc: the common situation (phase jumps)

Sample SP, one arc,, : after removing jumps, fixing the minimum to zero

Offset Ω is an arbitrary quantity: can we set it in some useful way?

A new set of observables: Phase slants leveled to Code

Operator <·> is a properly selected weighted (possibly robust) average

Build, arc by arc, the leveled slants SL

SL = SP - < SP – SC >

< SP – SC > = Ω - < m> - β - γ

SL = TEC + < m> + β + γ

Properties of SL

Noise is the same (neglected) of phase slants

Biased exactly as code slants

But: an arc dependent constant leveling error λ = < n> + < m> appears

Sample SC and SP with properly selected phase offset Ω = SL

* Evidence that calibration is needed: TEC is a positive quantity

*

One day, SC and SL arcs

Summary of the observables

SP = TEC + Ω

SC = TEC + m + β + γ

SL = TEC + λArc + β + γ

Ω Offset, constant but arbitrarily changing from arc to arc

β , γ Hardware biases: delays in electronics of transmitter and receiver. One β for satellite, one γ per station.

m Multi-path,

λ Leveling error, <m> , changing generally (but not arbitrarily) from arc to arc.

TEC The quantity to estimate, variable from observation to observation

Following topics will be discussed in the following

GPS ionospheric observables

Reliability of leveled slants

Problems with multipath

Problems with receivers?

TEC expansion

Reliability of the thin shell approximation

Calibration

The thin-shell, single-station, multi-day solution

of individual arc offsets

Validation

Use of ionospheric models to validate the calibration techniques

Features of observations, Code slants

SC = TEC + n + m + β + γ

Advantages: the electronic delays are physical quantities, stable or undergoing slow aging in controlled environmental conditions: they are generally considered constants over long times (up to 1 month).

One β per satellite, one γ for station: a favorable unknowns/observations budget.

n: strong radio noise (non linear techniques used to evaluate pseudo-ranges), but still a stochastic variable with zero mean (resulting in consistent estimations)

Can multipath m be considered a disturbance?

How to distinguish it from noise? Period of GPS orbits is 12 sidereal hours: day after day the same satellite will occupy the same position with an advance of ≈ 4 minutes: if same environment day after day, m will advance by the same amount.

Plot a fraction of arc of the same satellite day by day with an advance of ≈ 4 minutesNote: to avoid TEC variability, what is plotted for each arc is TEC(t) – TEC(t0), t0 being the beginning of each arc. Both SG and SΦ relative to the same arc are plotted .

Features of observations: Phase slants

SP = TEC + Ω

No significant noise and multipath (above slide)

Modest equations/unknown budget: one unknown per arc

Global single day solution, 200 stations

Unknowns: coefficients of TEC expansion plus around 1000 unknown offsets, compared to 200+30 hardware biases.

Possibility to use first differences (in time) of the observations of one arc. Only TEC coefficients remain: calibration relies entirely on the model used for the expansion.

Other possibility: solving by geodetic techniques for the ambiguities and therefore for the offsets.

Leveled slants: SL = TEC + λ + β + γ

λ = < m>

As for code slants, one unknown per satellite β and for station γ

Same observations/unknown budget of phase slants SP, apart the leveling error, constant arc by arc

Commonly assumed: disregard leveling error λ = < m>

In leveling error, the mean of a stochastic variable , <n> has been neglected as a quantity with (likely) zero mean: it can be considered a disturbance thatwill not significantly affect the ultimate accuracy of calibration.

Does the same holds for <m> ?

No: multi-path is not a stochastic variable and it has no zero mean

The close stations experiment can evidence this statement

Availability of close stations

Many co-located IGS stations are available:

darr/darw, dav1/davr, gode/godz, gol2/gold,kou1/kour, mad2/madr, mat1/mate

ohi2/ohi3, reyk/reyz, tcms/tnml, thu2/thu3, tid1/tid2, tid1/tidb, tid2/tidb, zimj/zimz

and the combinations of wtza, wtzj, wtzr, wtzt.

Besides IGS stations, a special set of observation has been set up by the group of La Plata University, Argentina (C.Brunini, F.Azpiliqueta).

Close to the IGS station “lpgs”, the additional stations “blue”, "red0" and "asht" have been set up for present investigation, whose characteristics will be described in (*).

Duration: days 182/205 and 262/269 , 2005

(*) Journal of Geodesy

DOI 10.1007/s00190-006-0093-1

Calibration Errors on Experimental Slant Total Electron Content (TEC) Determined with GPSL. Ciraolo, F. Azpilicueta, C. Brunini, A. Meza, S. M. Radicella

Updated availability of close station (2008)

cagl/cagz; cont/conz; darr/darw; dav1/davr; gode/godz;

gol2/gold; harb/hrao; hers/hert; irkj/irkm; irkj/irkt;

irkm/irkt; joz2/joze; kir0/kiru; lhas/lhaz; mad2/madr;

mat1/mate; mdvj/mdvo; mets/metz; mobj/mobn; nya1/nyal;

ohi2/ohi3; suth/sutm; tcms/tnml; thu2/thu3; tid1/tid2;

tid1/tidb; tid2/tidb; tixi/tixj; tro1/trom; tsk2/tskb;

usn3/usno; wtza/wtzj; wtza/wtzr; wtza/wtzs; wtza/wtzz;

wtzj/wtzr; wtzj/wtzs; wtzj/wtzz; wtzr/wtzs; wtzr/wtzz;

wtzs/wtzz; yakt/yakz; yar2/yarr; zimj/zimm;

Station 1

S1PRN = TEC + λ1 + β PRN+ γ1

The close stations experiment

< 100 m

TEC

β

S1 – S2 =γ1 - γ2 + λ1 − λ2

Not dependent on PRN

Station 2

S2PRN = TEC + λ2 + β PRN+ γ2

The close stations experiment

In equations of observation

S = TEC + β + γ + λConsider observations to satellite i from stations j e k

Sij = TECij + βi + γj + λArc__i

Sik = TECik + βi + γk + λArc_k

For close stations (up to few km) TECij = TECik satellite bias contribution is canceled

Sij - Sik = γj − γk + λArc_i – λArc_k

If contribution of leveling error is not significant, plotting Sij – Sik one gets points close to the difference γj − γk , a constant quantity for the investigated pair of stations.

Si1 – Si2 , i=1..all satellites, TECu

The situation for gol2/gold is rather uncommon

Most of times the situation is quite different as

a significant spread among satellites appears

As shown in following slides

Possible cause

the leveling error λ = < m > ?

SL1 – SL2 , all satellites

SP, zimj Arcs leveled to 0 minimum value

SP, zimm Arcs leveled to 0 minimum value

SP (zimj) - SP (zimm)

SL1 – SL2 , all satellites

SL1 – SL2 , all satellites

Is this spread due to multipath?The spread among satellites, according to

Sij - Sik = γj − γk + λArc_i – λArc_k

provides with an estimation of the spread of λArc_i – λArc_k around γj − γk

The split antenna experiment seems to confirm it.The receivers of "blue" and "red0", of the same firm, have been fed from the same antenna.

Implications: "blue" and "red0" see exactly the same multipath.

Besides IGS stations, a special set of observation has been set up by the group of La Plata University, Argentina (C.Brunini, F.Azpiliqueta).

Close to the IGS station “lpgs”, the additional stations “blue”, "red0" and "asht“have been set up to perform the following experiments

Close stations: different multipath; same or different way of processing multipath

Split antenna, receivers of same firm: same multipath, same way of processing it

Split antenna, receivers of different firms: same multipath, different way of processing

lpgsblue

red0

Same firm

asht

Different firm

Different multi-pathSame multi-path

< 100 m

Split antenna, same multipath, same type of receiver

Split antenna, same multipath, same type of receiver

To reduce errors in observations, what is needed is

Recipes to reduce multipath effects

-care antenna environment and radio-technical coupling

In the normal situation, the observed discrepancies amount to several TECu.

If this is due to multi-path only, great care must be taken in selecting a weighted average <·> using small weights when multi-path is expected to be large:

-avoid short arcs-care the selection of weights -use an elevation mask as higher as possible (where m is reasonably less strong)

empirically, using past experiencetrying to estimate them from the plots of SG – S, which according to the equations of the reported observables is m +n - <m + n>

W = 1 if Abs(SG – S)<Sigma SG – SW = Sigma SG – S / Abs(SG – S)2n

But are we dealing with actual multipath only?

For some station pairs, strange patterns appear.

In the following, station "wtzj" compared to the colocated "wtza", "wtzr", "wtzt", "wtzz", exhibits a strange pattern.

The problem is limited to "wtzj" , as the plots for other pairs are "normal".

Is it a thermal drift of station bias?

What will it happen to the calibration with discrepancies amounting to almost 25 TECu , and having no knowledge of the behavior of the station (evidenced only by the availability of close stations) ?

SL1 – SL2 , all satellites

SL1 – SL2 , all satellites

SL1 – SL2 , all satellites

SL1 – SL2 , all satellites

SL1 – SL2 , all satellites

SL1 – SL2 , all satellites

SL1 – SL2 , all satellites

Arcs leveled to 0 minimum value

Arcs leveled to 0 minimum value

SP (wtzj) - SP (wtzr)

Still: only multipath or some other problem?

Back to the split antenna experiment,

but using receivers of different firms.

Spread will appear again, suggesting that its cause is more the way by which multipath is processed rather than multipath itself.

lpgsblue

red0

Same firm

asht

Different firm

Different multi-pathSame multi-path

< 100 m

Split antenna, same multipath, different type of receiver

Split antenna, same multipath, different type of receiver

Conclusion of above experiments

Leveled to code slants are affected by the leveling error λ

The leveling error λ is most likely due to multipath (*)

Receivers of the same type produce similar λ’s, but there is no way to estimate their magnitude

Different types of receivers produce different λ’s observing the same ray

(*) other possible cause are possible, but not up to now investigated: studying scintillation it has been evidenced effect due to interference of other GPS satellites (still sidereal-time synchronous effects)

Is it correct modeling leveled slants SL disregarding λ?

For many station pairs, answer is negative

Still: no a priori method exists to notice that something is wrong unless availing two or more stations (see above plot of slants from close stations).

The results of the close stations experiment seem to evidence the need to introduce an additional satellite “bias”, the leveling error λ , dependent on the receiving station

(and the receiver type == way of extracting pseudorange).

Leveling error λ is an arc dependent unknown: this implies that

No advantage is taken using leveled slants SL with respect to phase slants (but this will need introducing one unknown per arc).

The choice of the calibration method

Aiming to

a simple solution (thin shell)

avoiding the problems of slants leveled to code SL (when leveling error is disregarded )

mitigating the errors of mapping function

It is natural to select a single station solution using phase slants SP or leveled slants SL

Notes about VEq approach

It takes automatically into account of plasmaspheric contribution

It is easier to model at low latitudes than actual vertical TEC

It presents some more difficulty to model at low elevations

The single station solution: Calibration

Observations

Phase slants SP

Assumptions

One thin shell at 400 km

Elevation mask: 10o

TEC expressed through VEq at the ionospheric point, by the mapping function sec χ

VEq expressed as a proper expansion of horizontal coordinates l, f with one set of coefficients at each time VEq(l, f) = Σncnpn(l,f)

Sijt = Σnc (t)n pn ( lijt , fijt ) sec χijt+ ΩArc

The unknowns are now the coefficients cn(t) and the offsets ΩArc

To solve the system

Sijt = Σnc (t)n pn ( lijt , fijt ) sec χijt+ ΩArc

extra assumptions are taken to reduce the number of coefficients Σnc (t)n

Using as horizontal coordinates Modified Dip Angle and Local Time, we can assume that for a set of adjacent epochs (up to ±15 minutes), the coefficients cn

(t) keep constant.

This allows also reducing computing resources during solution using commonly used standard methods for sparse systems.

After the solution of the system, we avail with

Calibrated slants along the observed rays TECijt = Sijt - ΩArc

“Mapped slants” at given coordinates lijt , fijt

Vertical TEC above the station (ionospheric point at the its zenith)

( ) ijtZenith

ijtZenithijt

nn

tn flpctVTec χsec,)( )(∑=

Performance of the proposed calibration method must be now investigated

1) A first look: will it provide same TEC’s from colocated stations?

2) Internal consistency: compute the residuals

Rijt = Sijt - Σnc (t)n pn ( lijt , fijt ) sec χijt- ΩArc

Small residuals mean good internal consistency, but do not help in asserting the accuracy of the method.

3) External consistency, namely the comparison with completely independent observations, should be the only way to assert the accuracy. Possible observations: Incoherent Scatter Radar (ISR), Two-Frequency Radar Altimeter (RA-2). Problems: very few ISR’s, RA-2 needs its own calibration.

Only possibility: using artificial truth data obtained using ionospheric models

A first look: worth adopting the above procedure for calibration?

Close station plots for wtza, wtzj, wtzr suggest that something is wrong with wtzj. Try arc offsets and standard biases calibration for the above stations

Standard solution, SL, β+γ

Proposed solution, SP, ΩArc

Proposed solution, SP, ΩArc

Proposed solution, SP, ΩArc

How do traditional and proposed solution compare?

In the following slides it can be seen that the two solutions agree in the average, but the difference in bias can amount to 10 TECu

The pattern of the jumps, similar for different satellites, simply indicates that something has changed in the receiver

Next topic: how can artifical data help in estimating

the reliability of calibration techniques?

How accuracy of calibration techniques can be estimatedExamination of residuals

Resijt = Sijt - Σnc tn pn ( lijt , fijt ) sec χijt- ΩArc

After a calibration run will provide with useful information about the

Internal consistency of the solution

Residuals are plotted in the following examples for few sample stations.

Standard deviation of the individual samples is reported.

Internal consistency of the method is estimated from the residuals (actual data)

Resijt = Sijt - Σnc tn pn ( lijt , fijt ) sec χijt- ΩArc

Residuals, actual data

Residuals, actual data

Residuals, actual data

Sigma of the shown sample residuals ranges from ≈ .5 to 4 TECu according to latitude.

Is this an estimation of the accuracy of the calibration?

No, as this requires a comparison with truth data, which are unavailable

(Incoherent Scatter Radar, Radar Altimeter may help, but are not sufficient).

What can look more like truth data?

Artificial data produced by Ionospheric Models.

But keeping in mind that agreement with artificial data is a condition

necessary but not sufficient to validate the method

The artificial data

Ionospheric models enable to estimate median electron density at some time at some geographic location, i.e. given date and time, latitude, longitude, height.

Ne = Ne(t,φ,λ,h)TEC is the integral of electron density along the ray-path from satellite to receiver,

which will be numerically evaluated as the sum

or with any more effective numerical algorithm (Gauss, …)

∫= dsPNTEC e )(

iie s)δ(PNTEC ∑≈

iiee s)δ(PNdsPNTEC ∑∫ ≈= )(

Pi , dsiRx

Sat

Pi, point on the generic ith shell

δsi increment in arc length

Model TEC computation

Divide the path in elements δsi

At each point Pi compute the electron density Ne(Pi)provided by the model

Multiply by the element length δsi

Cumulate all elements

Simple uses of artificial data: the mapping function

Which errors do affect the standard approach (actual vertical TEC) of mapping function?

Using an artificial ionosphere:

Compute χ

Compute Slant S

Compute Vertical TEC V at the Ionospheric Point

Error: S – V sec χ

Plot Error distribution

Station

Vertical

To GPS

ds

dh

h

χP

Ionosphere

Simple uses of artificial data: VEC and VEqIn the Single-Station / Arc Offset calibration the Vertical Equivalent TEC VEq for which it is exactly S = VEq sec χ is used.

How different is VEq from actual Vertical TEC (VEC) ?

Using an artificial ionosphere:

Compute χ

Compute Slant S

By definition VEq = S cos χ

Compute Vertical TEC V at the Ionospheric Point VEC

Plot VEC, VEq

Plasmasphere can be included too using a suitable model

PRx

Sat

hRef

VEC

Sχ

Integration paths for

S

VEC

VEq = S cos χ

Simple uses of artificial data: How much VEC and VEq differ?

Test of Single-Station, Arc-Offset solution

Generation of artificial truth data

Given all slants actually observed and archived

in a (quasi) complete set of IGS stations (≈ 200 per day) for year 2000 for days 88-91 ( March 28-31)

Re-compute them using NeQuick (Az =150), integrating up to 2000 km

Therefore:

Not only the actual GPS constellation has been preserved for the reference period, but also the possible lack of observations (this will affect the solution)

Internal consistency: Residuals, simulated data

Resijt = Sijt - Σnc tn pn ( lijt , fijt ) sec χijt- ΩArc

Set of slants from IGS

Recompute using NeQuick

SOut - SIn

Arrange slants by arcsCorrect for phase jumpsLevel ArcEvaluate Arc OffsetsCompute SOut

Truth Data SIN

Testing the calibration procedure

SOut – SIn are plotted vs time

Worth (but expected) noting that errors at low latitudes are larger

Remark about highlighted arc:

errors show a weakness of the solution.

These errors occur for arcs of low elevation also if, in some case, of long duration.

Processing real data, there is no chance to know if the subject arc is ill-calibrated (unless in presence of very strong errors)

Testing the solution with simulated data will (likely) enable to find a more effective way of avoiding such errors, or in a last instance, rejecting them

SlantOut-SlantIn, TECu

SlantOut-SlantIn, TECu

SlantOut-SlantIn, TECu

SlantOut-SlantIn, TECu

An overall look to the errors: SOut – SIn, whole set

0.12% < -10 0.067 % > 10

An overall look to the errors: SOut – SIn, probability density

Error’s behavior vs latitude: percentiles, whole set

Simulation: role of multi-path contribution λAn arbitrary set of satellite + receiver biases + multipath errors is added to model slants

Station bias γ = 25

Satellite biases βi = 10 * (Rnd() - Rnd()) , i=1,..,32

LevelingError λArc = 10 * Rnd()

Arc Offset ΩArc = 1000 * Rnd()NextData are processed both by traditional and arc offset single-station

calibration.

Arc =1.. Number of Arcs

Set of slants from IGS

Recompute using NeQuick

SOut - SIn

Arrange slants by arcsCorrect for phase jumpsAdd biases β + γ + λLevel ArcsEvaluate Traditional/ Arc OffsetsCompute SOut , VEq

Truth Data SIN

( VEq )

Traditional, SOut - SIn

Traditional, VEq computed / VEq True

Arc Offset, SOUT - SIn

Arc Offset, VEq computed / VEq True

Thank you